What is Ox Alpha, really?

Ox Alpha is a reasoning model that showed up on OpenRouter under the masked "Stealth" provider. Its creators are not public — which is unusual for a model this capable — and it's free to use during the stealth preview.

The headline numbers

• 1,048,576 tokens of context (a full 1M)
• 131,072 tokens of maximum output
• Accepts text, image, and video as input; outputs text
• Reasoning is mandatory, with three effort levels: low, high, max
• $0 cost while in stealth

Why people care

Two reasons. First, the context window is enormous — you can put an entire codebase in a single prompt. Second, it's positioned squarely at coding and long-horizon agent work, which is exactly where a strong-reasoning, large-context model earns its keep.

What's genuinely unknown

The tokenizer is reported as "Other", the knowledge cutoff isn't disclosed, and the provider is masked. Treat benchmark claims you see floating around with healthy skepticism — including any on this site that we didn't test ourselves.

If you're evaluating it, test it on your tasks. A stealth model's marketing is the least reliable thing about it.